Researchers Nicolas Boizard and colleagues investigate whether distilling reasoning traces from strong teacher models is worth the computational cost. They conduct a controlled experiment where a single teacher model generates paired instruction fine-tuning (IFT) and reasoning outputs for identical prompts. The results show that at matched floating-point operations (FLOPs), IFT often lies on or near the Pareto frontier across various configurations, while reasoning reaches this frontier only on open-ended tasks at 7 billion parameters and above. Even in these cases, a sequential curriculum mixing a small proportion of reasoning data with IFT captures most of the accuracy benefit at significantly lower compute cost. --- Why it matters: This study matters to AI researchers because it challenges the conventional wisdom that distilling reasoning traces is necessary for building capable small language models. The findings suggest that this approach may not be worth the computational cost, and that alternative methods like instruction fine-tuning could achieve similar results with less resource-intensive training. Source: https://arxiv.org/abs/2509.22193